Why This Matters More Than Another Chatbot Wrapper
Every cross-border operator I know is drowning in the same paradox: we have more data than ever about our customers, our supply chain, and our ad spend, yet we make decisions with less context than ever. Your Amazon rep says one thing, your Shopify dashboard says another, your freight forwarder’s spreadsheet contradicts both, and somewhere in a Slack thread from three weeks ago is the actual answer to the question you’re about to ask. We’ve bought every SaaS tool that promises a “single source of truth,” and we’ve ended up with seventeen sources of partial truth. The product I’m about to dissect isn’t built for e-commerce specifically, but the underlying bet — that an AI agent should compile context upfront rather than bolt memory on as an afterthought — is the exact bet every serious seller needs to be making about their own operations. If you can teach an agent to resolve contradictions between a Slack thread and a Google Doc, you can teach it to resolve contradictions between a supplier invoice and a customs broker’s estimate. That’s the thesis. Everything else is implementation.
The Problem: Your Business Has Amnesia, and So Does Every AI Tool You’ve Tried
Let’s be brutally honest about the state of “AI assistants” in e-commerce operations. We’ve all tried the chat-based tools that promise to know our business. You connect your Shopify store, your Amazon Seller Central account, maybe your Klaviyo instance, and you ask a question like “What was our gross margin on the Germany launch last quarter?” What you get back is either a hallucinated number that sounds plausible, or a weaselly disclaimer that the tool “couldn’t find that information.” The reason is structural: these tools treat memory as a retrieval problem, not a compilation problem. They scrape your connected apps on demand, try to piece together context in real time, and fail the moment the answer requires synthesizing information from more than one source.
The founders of Almanac — Kushagra Chitkara and his two co-founders — describe hitting this exact wall themselves. As their launch post explains, they tried to set up an agent with company-wide context and found it was a pain: hosting it, building an OAuth app for every connector by hand, feeding it context manually, and constantly fighting its memory. Their YC batchmates were hitting the same wall, so they built the thing they wished existed. That’s the origin story of every good tool, but what’s interesting here is the architectural bet they made, not the origin story.
The core insight, as Chitkara frames it, is that “memory has to be compiled upfront with real compute, not bolted on as an afterthought.” That’s a direct jab at the entire category of retrieval-augmented generation (RAG) tools that have flooded the market — the ones that just stuff your documents into a vector database and hope for the best. Almanac’s approach is different: it compiles your connected tools into two structured wikis, a personal one and a company one, and it reads those wikis before doing anything. This isn’t semantic search; it’s active curation. The system decides whether new information is a new page, an edit to an existing page, or a contradiction that needs resolving.
For a cross-border seller, this distinction is everything. Your business runs on contradictions: the P&L says one thing, the bank statement says another, the inventory report says a third. A tool that can’t resolve those contradictions isn’t a brain, it’s a filing cabinet with a chat interface.
How Almanac Actually Works (And What It Gets Right)
The Wiki as a Compiled Brain
The most interesting architectural decision here is the two-wiki split. Almanac maintains a personal wiki (who you are, your preferences, your people) and a company wiki (what you’re building, the roadmap, the blockers). This matters for a reason that goes beyond privacy theater. In a typical e-commerce operation, the founder’s private context — relationships with suppliers, personal risk tolerance, which marketplace manager you trust and which one you don’t — genuinely informs business decisions, but it shouldn’t be visible to everyone on the team. The split lets the agent use private context to inform company tasks without exposing that private information, which is exactly the question one commenter asked about in the launch thread.
The second piece is what they call “dreaming” — nightly batch compute that consolidates the day’s affairs into permanent memory. When a commenter asked whether this runs as daily batch jobs on Modal or immediately when a new doc drops, Chitkara’s answer was revealing: “It runs nightly. The way we think about context is quite similar to how humans hold memory. You have your immediate context, which we maintain when new things drop. And dreaming is where we consolidate the day’s affairs to your permanent memory.” That’s a genuinely different architecture from the real-time-everything approach most AI tools take, and it’s arguably more aligned with how businesses actually operate. You don’t need real-time synthesis of every Slack message; you need a daily reconciliation of what actually matters.
The third piece is the agent’s own computer — a real browser and terminal. For tools with no API, it signs in and clicks around like a human would. This is the part that should make e-commerce operators sit up and pay attention, because the reality of our tooling stack is that half of it doesn’t have APIs worth using. Your freight forwarder’s portal, your overseas warehouse management system, your customs broker’s tracking page — these are all browser-based tools with no API access. An agent that can operate a browser is an agent that can actually work with the messy, un-API’d reality of cross-border operations.
The “It Just Knows Me” Feeling
The self-updating brain is what creates what Chitkara calls the “it just knows me” feeling. The agent compiles your connected tools into wikis, reads them before doing anything, and reviews every update against the existing wiki before it lands. This is the opposite of the approach taken by tools like Notion AI or Mem, which treat memory as a searchable archive. Almanac treats memory as a living document that must be actively maintained, contradictions resolved, and outdated information pruned.
For a seller running multiple marketplaces, this distinction matters. Your Amazon Seller Central data, your Shopify admin, your TikTok Shop backend, and your Etsy dashboard all tell slightly different versions of the same story. An agent that can compile those into a single coherent wiki — and flag the contradictions rather than papering over them — is genuinely useful. An agent that just searches across all of them on demand is marginally better than doing the search yourself.
### Why Amazon Sellers Should Care More Than Shopify Ones
Here’s my contrarian take: this tool is more relevant to Amazon FBA operators than to Shopify DTC brands, and the reason is structural. Shopify merchants live in a relatively clean data world. Your store, your app integrations, your marketing tools — most of them have solid APIs, and the data is reasonably well-structured. Amazon sellers live in a world of legacy systems, half-broken APIs, and information that lives in Seller Central, in rep emails, in FBA inventory reports, and in third-party tools like Helium 10 or Jungle Scout that don’t always play nice with each other.
The browser automation piece is the killer feature for Amazon sellers. When an agent can sign into Seller Central and click around like a human — checking your account health, reviewing your FBA inbound shipments, monitoring your Buy Box percentage — it’s doing something that no API-based tool can do. And when that browser interaction gets compiled into a wiki that the agent reads before making decisions, you start to get a tool that actually understands the state of your Amazon business, not just the data.
The personal vs. company wiki split also matters more for Amazon sellers, because the person who runs the Amazon channel is often the founder or a single operator, and their private context — relationships with suppliers, knowledge of which products are being phased out, awareness of pending IP complaints — genuinely changes how they’d want tasks executed. A tool that respects that boundary while still using the context to inform company tasks is rare.
What Cross-Border Sellers Can Actually Borrow From This
The Connector Strategy: Start With What You Actually Use
Almanac’s connector list is telling: Gmail, Calendar, Granola, PostHog, GitHub, and “more.” It’s not trying to integrate with every tool on earth. It’s starting with the tools that generate the most context — email, calendar, and the tools where work actually gets discussed. For a cross-border seller, the equivalent list would be Gmail, your marketplace seller central accounts, your freight forwarder’s portal, and your accounting software. The lesson is to start with the tools that hold the most unstructured context, not the ones with the cleanest APIs.
The “Dreaming” Pattern: Nightly Reconciliation
The nightly batch compute pattern is something every seller should steal, even without the tool. Set aside 30 minutes at the end of each day to reconcile what happened: which SKUs moved, which ads converted, which suppliers responded, which marketplace policies changed. The act of compiling that into a coherent daily summary — a “wiki” of your operations — is what separates operators who understand their business from operators who just react to it. The tool automates this; you can start by doing it manually.
### Where the Math Breaks
Let’s talk about the elephant in the room: the permission model. When a commenter asked how to control which tools Almanac has access to — specifically wanting read-only access to email — Chitkara’s response was blunt: “Currently we ask for all the permissions upfront.” That’s a dealbreaker for a lot of e-commerce operators, and it should be. Granting an AI agent full read/write access to your Gmail, your marketplace accounts, and your payment systems is a massive security exposure. One bad prompt, one prompt injection attack through a supplier email, and the agent could do real damage.
The launch thread shows other gaps too. When asked how the system handles conflicting context — an old Slack thread versus a newer doc — Chitkara’s answer was that the “dreaming” process resolves these conflicts before making them permanent. That’s the theory. In practice, conflict resolution is the hardest problem in knowledge management, and the answer “we spend compute upfront to resolve it” doesn’t fully explain how the system decides which source is authoritative. Is it recency? Source priority? Semantic consistency with other wiki pages? The post doesn’t say.
There’s also the question of cost. “Sleep time compute” is a nice phrase, but nightly batch processing of your entire company’s data across all connected tools is not free. The pricing page mentions a 7-day trial on every plan, but actual pricing isn’t disclosed in the launch post. For a solo seller or a small team, the cost of running this kind of continuous compilation could easily exceed the value it delivers — especially when the alternative is a weekly manual review of your key metrics.
The Competitive Landscape: What Almanac Is Up Against
The AI-knowledge-management space is crowded, and Almanac’s differentiation is real but not unassailable. The most direct comparison is Glean, which also tries to compile company knowledge into a searchable index. Glean’s approach is more enterprise-focused, with a heavier emphasis on security and compliance, but it doesn’t have the agentic execution piece — Glean tells you where the answer is; Almanac tries to do the work. That’s a meaningful difference for e-commerce operators who don’t just need answers, they need tasks completed.
On the agentic side, tools like Mastra and CrewAI let you build custom agents, but they require significant engineering effort. Almanac’s pitch is that you don’t need to build anything — you connect your tools, and the agent figures out how to use them. That’s compelling for a seller who doesn’t have a dedicated engineering team, which is most sellers.
The closest comparison might be Sana, which also tries to compile company knowledge and make it actionable. But Sana’s focus is more on internal knowledge sharing and onboarding, while Almanac’s focus is on task execution. The difference is subtle but important: Sana helps your team find information; Almanac tries to act on it.
Where Almanac falls short is in the e-commerce-specific integrations. The launch post mentions Gmail, Calendar, Granola, PostHog, and GitHub — all developer or general business tools. There’s no Shopify connector, no Amazon Seller Central integration, no Klaviyo connector. For a cross-border seller, that means the tool would require significant manual setup to be useful: you’d need to connect your Gmail (where your supplier communications live), your calendar (where your meetings with freight forwarders and marketplace managers live), and then manually feed in the data from your marketplace dashboards. That’s a lot of upfront work for a tool that’s supposed to save you time.
What I’d Watch / Test Next
Here’s what I’d do this week if I were running a cross-border operation and wanted to test the Almanac thesis without going all-in:
1. Start with the wiki discipline, not the tool. Before you connect anything, write down the 20 facts about your business that matter most: your top 5 SKUs by margin, your current inventory levels in each warehouse, your open supplier orders, your pending marketplace policy changes. That’s your company wiki. Then write down the 10 personal facts that inform your decisions: your risk tolerance, your key supplier relationships, your preferred communication channels. That’s your personal wiki. If you can’t write these down in an hour, you have a knowledge problem that no AI tool will fix.
2. Test the “dreaming” pattern manually. For the next 7 days, spend 15 minutes at the end of each day writing a daily reconciliation: what changed, what contradicted what you thought you knew, what needs follow-up. At the end of the week, read all seven entries and see what patterns emerge. That’s what Almanac is doing with compute; you can do it with a text file.
3. If you want to try the actual tool, connect only your Gmail and calendar first. Don’t grant access to anything sensitive. The 7-day trial is long enough to see whether the wiki compilation actually produces insights you didn’t already have. Ask it the questions you’d normally ask your team: “What’s blocking the Germany launch?” “Who haven’t we followed up with on the supplier audit?” “What changed in our ad spend this week?” If the answers are genuinely useful, consider expanding access. If they’re generic, you’ve learned something about the limits of compiled memory.
4. Watch the permission model. The “all permissions upfront” approach is fine for a trial, but if you’re considering production use, wait until they offer granular access controls. The launch thread shows this is a known gap, and it’s the kind of thing that gets fixed quickly when enough users ask for it.
The bigger bet here isn’t Almanac specifically — it’s the architectural principle that memory must be compiled, not retrieved. Whether or not this particular tool survives, that principle is going to shape every AI tool you use in the next three years. Start building the discipline now, and you’ll be ahead of the curve when the tools catch up to the reality of cross-border operations.
